Content restriction in generative AI manifests as explicit or implicit censorship. Explicit censorship uses predefined rules to block content like hate speech or illegal material, employing keyword blacklists, pattern-matching, or classifiers (Gillespie 2018). DeepSeek’s models, aligned with Chinese regulations, use real-time filters to block politically sensitive content, such as government criticism, ensuring legal compliance (WIRED 2025). When triggered, these filters issue refusals, enforcing standards bluntly. Implicit censorship arises from biases in training data or fine-tuning, suppressing perspectives via skewed output probabilities. Reinforcement Learning from Human Feedback (RLHF), where raters rank responses for qualities like harmlessness, embeds normative assumptions (Christiano et al. 2017). Grok 3, despite its “truth-seeking” branding, reportedly avoids unflattering mentions of certain figures, suggesting implicit filtering (TechCrunch 2025). Such biases, also in models like GPT-3, marginalise minority viewpoints (Bender et al. 2021).
Censorship splits into safety guardrails and ideological filtering. Safety guardrails mitigate harm, restricting content like misinformation or illegal material per regulations like the EU’s Digital Services Act. DeepSeek’s censorship of state-opposed content exemplifies guardrails tailored to legal mandates, though it spans broad political topics (WIRED 2025). Ideological filtering suppresses content based on political or cultural preferences, often opaquely. Fine-tuning on curated datasets embeds biases, as when models prioritise dominant narratives (Gururangan et al. 2020). Grok 3’s selective refusals, avoiding criticism of specific individuals, have sparked ideological bias claims, despite reduced censorship goals (TechCrunch 2025). Efforts like Constitutional AI, which aligns models to predefined principles, aim to enhance transparency in addressing these biases but risk entrenching the developers’ values, perpetuating normative constraints on outputs (Bai et al. 2022).
Alignment, the process of tuning AI systems to reflect desired values, is central to censorship. RLHF aligns models to avoid harmful or sensitive content, but raters’ subjective preferences can embed ideological biases, leading to implicit censorship (Christiano et al. 2017). For example, DeepSeek’s alignment enforces strict regulatory compliance, filtering politically sensitive topics, while Grok 3’s alignment aims for minimal censorship but still exhibits selective biases (WIRED, 2025; TechCrunch, 2025). Misalignment or “alignment faking,” where models superficially comply but retain biased behaviours, complicates censorship control (Anthropic, 2024). Realignment can alter censorship by modifying model behaviour post-training. Perplexity’s R1 1776 model, derived from DeepSeek’s R1, was fine-tuned using a 40,000-prompt multilingual dataset targeting 300 topics censored by Chinese regulations, such as political dissent. A censorship classifier identified restricted prompts, and fine-tuning with Nvidia’s NeMo 2.0 framework adjusted model weights to bypass DeepSeek’s regulatory filters, enabling factual responses on sensitive issues (Perplexity AI, 2025).
In conclusion, censorship in generative AI is a multifaceted phenomenon, enacted through a combination of explicit technical guardrails and subtle, implicit biases embedded via alignment processes. The distinction between mitigating genuine harm and enforcing a particular ideology remains a persistent and often opaque challenge, managed by a complex socio-technical system susceptible to bias at every stage, from data collection to human moderation. The very malleability of these systems, demonstrated by the potential for both 'alignment faking' and deliberate 'realignment', underscores a critical reality: AI-driven content restriction is not a static technical problem but a continuous series of normative choices. This highlights the urgent need for transparent, accountable governance frameworks that can navigate the inherent tensions between ensuring safety and preserving expressive freedom in the evolving information ecosystem.
References:
1. Anthropic. 2024. ‘Alignment Faking in Large Language Models’. Anthropic Research Reports. https://www.anthropic.com/research/alignment-faking ^ Back
2. Bai, Yuntao, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen et al. 2022. ‘Constitutional AI: Harmlessness from AI Feedback’. arXiv preprint arXiv:2212.08073. https://arxiv.org/abs/2212.08073 ^ Back
3. Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021, March. ‘On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?’. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 610–623. ^ Back
4. Christiano, Paul F., Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. ‘Deep Reinforcement Learning from Human Preferences’. Advances in Neural Information Processing Systems 30. https://proceedings.neurips.cc/.../d5e2c0adad503c91f91df240d0cd4e49 ^ Back
5. Gillespie, Tarleton. 2018. Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions That Shape Social Media. New Haven: Yale University Press. ^ Back
6. Gururangan, Suchin, Ana Marasović, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. ‘Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks’. arXiv preprint arXiv:2004.10964. https://arxiv.org/abs/2004.10964 ^ Back
7. Perplexity AI. 2025. ‘Open-sourcing R1 1776’. Perplexity Blog. https://www.perplexity.ai/hu/hub/blog/open-sourcing-r1-1776 ^ Back
8. TechCrunch. 2025. ‘Grok 3 Appears to Have Briefly Censored Unflattering Mentions of Musk and Trump’. https://techcrunch.com/2025/02/23/grok-3-appears-to-have-briefly-censored-unflattering-mentions-of-trump-and-musk/ ^ Back
9. WIRED. 2025. ‘Here’s How DeepSeek Censorship Actually Works—And How to Get Around It’. https://www.wired.com/story/deepseek-censorship/ ^ Back
